README.md

metadata

title: My Face Enhancement Space
emoji: 😊
colorFrom: blue
colorTo: green
sdk: gradio
sdk_version: 5.25.2
app_file: demo.py
pinned: false

Face Enhance

A tool for improving facial consistency and quality in AI-generated images. Dramatically enhance facial fidelity while preserving the original image's background, lighting, and composition.

Installation

Prerequisites

• Python 3.11 or higher
• 1 GPU with 48GB VRAM
• At least 50GB of free disk space

Setup

1. Set up your Hugging Face token:

   • Create a token at Hugging Face
   • Log into Hugging Face and accept their terms of service to download Flux.1-dev
   • Set the following environment variables:

     export HUGGINGFACE_TOKEN=your_token_here
     export HF_HOME=/path/to/your/huggingface_cache
     

   • Models will be downloaded to $HF_HOME and symlinked to ./ComfyUI/models/

2. Create the virtual environment:

   python -m venv venv
   source venv/bin/activate
   python -m pip install -r requirements.txt
   

3. Run the install script:

   python install.py
   

   This will

   • Install ComfyUI, custom nodes, and remaining dependencies to your venv
   • Download all required models (Flux.1-dev, ControlNet, text encoders, PuLID, and more)

4. Run inference on one example:

   python test.py --input examples/dany_gpt_1.png --ref examples/dany_face.jpg --out examples/dany_enhanced.png
   

   Arguments

   • --input (str): Path to the input image.
   • --ref (str): Path to the reference face image.
   • --output (str): Path to save the output image.
   • --crop (store_true): Flag to face crop the reference image. Default: False.
   • --upscale (store_true): Flag to upscale the (cropped) reference image. Default: False.
   • --caption (store_true): Flag to caption the input image. Default: False.
   • --id_weight (float): Face ID weight. Default: 0.75.

Gradio Demo

A simple web interface for the face enhancement workflow.

1. Run python demo.py

2. Go to http://localhost:7860. You may need to enable port forwarding.

Running on ComfyUI

Using the ComfyUI workflows is the fastest way to get started. Run python run_comfy.py. There are two workflows:

• ./workflows/FaceEnhancementProd.json for face enhancement
• ./workflows/FaceEmbedDist.json for computing the face embedding distance

Notes

• The script and demo run a ComfyUI server ephemerally
• Gradio demo is faster than the script because the models remain loaded in memory and ComfyUI server is booted up.
• All images are saved in ./ComfyUI/input/scratch/
• FaceEnhancementProd.py was created with the ComfyUI-to-Python-Extension and re-engineered for efficiency and function.
• Face cropping, upscaling, and captioning are unavailable; these will be added in an update.

Troubleshooting

• Out of memory errors: If your GPU has less than 48 GB VRAM, install Flux.1-dev at fp8 precision.
• Face detection issues: This method works for photorealistic images of people. It may not work on cartoons, anime characters, or non-human subjects.
• Downloading models fails: Check your Hugging Face token has proper permissions.